Zobraziť v EDS

MCNet: a multi-source remote sensing image cascade network for accurate and efficient tailings pond segmentation.

Uložené v:
Podrobná bibliografia
Názov: MCNet: a multi-source remote sensing image cascade network for accurate and efficient tailings pond segmentation.
Autori: Wang, Pan1 (AUTHOR) wangpan@student.cumtb.edu.cn, Zhao, Hengqian1,2,3 (AUTHOR) zhaohq@cumtb.edu.cn, Liu, Zhiguo1 (AUTHOR), Xu, Fei1 (AUTHOR), Fu, Hancong1 (AUTHOR), Mao, Jihua1 (AUTHOR)
Zdroj: Advances in Space Research. Nov2025, Vol. 76 Issue 10, p5955-5972. 18p.
Predmety: *TAILINGS dams, *REMOTE sensing, *IMAGE processing, *ARTIFICIAL neural networks, *SATELLITE-based remote sensing
Geografický termín: CHINA, HEBEI Sheng (China)
Abstrakt: • Proposed a new approach called MCNet for tailings pond segmentation. • Reduced interference from non-tailings pond scenes through cascade network. • Integrated Sentinel-2 and Jilin-1 to enhance accuracy and processing speed. • Provided a scalable solution for large-scale remote sensing monitoring. Efficient and accurate extraction of tailings ponds from large-scale remote sensing images is essential for effective safety monitoring. However, existing methods are inefficient and susceptible to complex background interference. To address these limitations, a multi-source remote sensing image cascade network (MCNet) was designed to quickly and accurately extract tailings ponds from large scene remote sensing images. The model synergistically couples low-resolution multispectral Sentinel-2 images and high-resolution Jilin-1 images through a two-stage cascade framework. The model first quickly identifies scenes containing tailings ponds based on ResNet50 model and Sentinel-2 remotely sensed images, and then processes the reserved areas on the Jilin-1 images through a semantic segmentation model to extract details about the tailings ponds. To evaluate the effectiveness of the model, experiments were conducted in Chengde City, Hebei Province, China, focusing on tailings pond extraction. The results demonstrate that the MCNet model has the highest segmentation performance, with an Intersection over Union (IOU) reaching 91.56 %, and an inference speed that is approximately five times faster than DeepLabv3 model. Additionally, using Sentinel-2 false color images to classify tailings ponds is better than that of true color images, highlighting the importance of spectral diversity for tailings pond identification. Finally, the tailings pond in Chengde City was extracted based on the MCNet model. The results showed that the area of the tailings pond in the city was 165.377 km2, and the extraction time was only 0.237 h. This approach provides a scalable solution for large-scale remote sensing monitoring. [ABSTRACT FROM AUTHOR]
Databáza: Academic Search Index
Popis
Abstrakt:• Proposed a new approach called MCNet for tailings pond segmentation. • Reduced interference from non-tailings pond scenes through cascade network. • Integrated Sentinel-2 and Jilin-1 to enhance accuracy and processing speed. • Provided a scalable solution for large-scale remote sensing monitoring. Efficient and accurate extraction of tailings ponds from large-scale remote sensing images is essential for effective safety monitoring. However, existing methods are inefficient and susceptible to complex background interference. To address these limitations, a multi-source remote sensing image cascade network (MCNet) was designed to quickly and accurately extract tailings ponds from large scene remote sensing images. The model synergistically couples low-resolution multispectral Sentinel-2 images and high-resolution Jilin-1 images through a two-stage cascade framework. The model first quickly identifies scenes containing tailings ponds based on ResNet50 model and Sentinel-2 remotely sensed images, and then processes the reserved areas on the Jilin-1 images through a semantic segmentation model to extract details about the tailings ponds. To evaluate the effectiveness of the model, experiments were conducted in Chengde City, Hebei Province, China, focusing on tailings pond extraction. The results demonstrate that the MCNet model has the highest segmentation performance, with an Intersection over Union (IOU) reaching 91.56 %, and an inference speed that is approximately five times faster than DeepLabv3 model. Additionally, using Sentinel-2 false color images to classify tailings ponds is better than that of true color images, highlighting the importance of spectral diversity for tailings pond identification. Finally, the tailings pond in Chengde City was extracted based on the MCNet model. The results showed that the area of the tailings pond in the city was 165.377 km2, and the extraction time was only 0.237 h. This approach provides a scalable solution for large-scale remote sensing monitoring. [ABSTRACT FROM AUTHOR]
ISSN:02731177
DOI:10.1016/j.asr.2025.08.047